Report Open Access
A method for the pseudonymization of speech is presented that allows to obfuscate the identity of a speaker in untranscribed running speech. The approach is to manipulate the spectro- temporal structure of the speech to simulate a different length and structure of the vocal tract, as well as a different pitch and speaking rate. The method is deterministic, and partially re- versible. The extend of the changes is adjustable and gradual. A series of ABX listening experiments show that both experts and non-experts identify speakers in less than 70% of forced choice pairs while listeners are able to identify over 90% of speak- ers without pseudonymization. Reverting the procedure, de- pseudonymization, is partially effective. Some pseudonymiza- tion targets, e.g., those simulating a long vowel tract, are more amenable to de-pseudonymization than others. The method also works differently on female and male voices. Depending on the pseudonymization target, female speakers were less well iden- tified after pseudonymization and de-pseudonymization than male speakers.